[erikwb]

The J713 refused system sleep in four places: the DCP (the retained T8132 session had "no qualified system-sleep handshake"), the retained display and radio DARTs, the radio's PCIe port, and genpd gating the GPU's GFX domain under the running G16 firmware. With this series it sleeps from systemctl suspend and from the lid, and wakes on the keyboard, trackpad, power button or lid, with display, GPU, Wi-Fi and audio intact.

The display handshake

On IOMFB 26.x the handshake is setPowerState (A473) on the out-of-band channel, which the T8132 words carry as stream 1 (bit 10).

• Off: with the panel blanked, setPowerState(0) makes the firmware log power on -> off, disable its M3 helpers and switch to run mode 0.
• On: setPowerState(1) brings it back (run mode 0 → 4, M3 firmware reloaded), and the next present scans out.
• Callbacks: the firmware calls back about ten times during each transition. They're answered with exact shapes.

Patches 1-3 add stream-1 calls to the transport (calls nested in a stream-1 callback stay on stream 1), the callback shapes and answers, and the suspend/resume hooks. apple-drm is a device-link consumer of the DCP, so it has blanked before the DCP suspends and restores the CRTC after it resumes. The DCP keeps its domains on through the wakeup path, as the older DCPs do.

Retained DMA owners

The radio's PCIe port and DART and the display DARTs have no power domain of their own in the tree, and the DCP holds the display domains on. So suspend-to-idle leaves their live state in place, and patches 4-5 let them through for s2idle only; every other sleep state is still refused. The DARTs save their registers only so resume can check that nothing moved.

If something did move, the consumers' DMA is no longer what they set up, and a resume error alone doesn't stop the PM core from resuming them. So before failing resume, the DART quarantines them through the same paths as a translation fault:

• the managed radio streams are marked failed and their DMA is disabled, so later maps are refused;
• the borrowed display consumers (the DCP and its frontends) are marked failed, so their later maps are refused.

GPU

asahi-g16 has no system-sleep callbacks, so genpd's noirq power-off gated GFX, and with it the TVM gate, under the live firmware. The power-on at resume then timed out (GFX target=0xf … error=-110). A pmgr-pwrstate domain marked apple,sleep-retained declines that power-off while the system sleeps (patches 6-7). Runtime PM is unchanged.

Wakeups

Every ASC mailbox receive IRQ is IRQF_NO_SUSPEND. The G16 firmware's idle ring-service kicks (~1.9/s) and an AOP report (1/s) woke the CPUs about three times a second:

	Sleep	Wakeups	Deep idle
Before patches 8-10	938 s	1472	512 s
After	148 s	3	146.9 s

A mailbox marked apple,hold-rx-in-sleep masks its receive IRQ from noirq suspend to noirq resume (patches 8-9). The firmware throttles once its outbox is full: after 148 s, 11 GPU and 3 AOP messages were waiting.

Patch 10 sets both flags for the GPU. The AOP mailbox flag is a one-line follow-up on the AOP node from #434:

&aop_mbox {
	status = "okay";
	apple,hold-rx-in-sleep;
};

Testing

• J713: this series plus arm64: keep J713 GPU/DCP non-coherent when CLIDR_EL1.LoC == 0 (fixes J713 graphics on 7.3) #429/cpuidle: apple: add T8132 and enable it on J713 (s2idle groundwork) #432/soc: apple: aop: T8132 (M4 MacBook Air) AOP and microphones #434/ASoC: apple: J713 (M4 MacBook Air) headphone jack on the AOP (depends on #434) #437/ASoC: apple: J713 (M4 MacBook Air) speakers on the AOP (depends on #437) #439 on 26.6.2 firmware.
• Real s2idle: systemctl suspend (up to ~16 min asleep) and lid close/open via logind. Display, GPU (glmark2 630-642, unchanged), Wi-Fi and the AOP audio all come back.
• Dry runs: pm_test=devices and pm_test=platform both pass.
• Host tests:

  • apple-dcp selftests: unchanged from integration/aurora (57/63, with the same 6 failures on the bare base). New cases cover stream-1 decoding and routing, and the refused present while powered off.

  • m4-dcp-errors passes, including its case where a failed display domain refuses maps.

  • bcm4377-radio-sids.py passes, with its negative controls.

  • New dart-s2idle-retention.py, on the extracted production state check, quarantine and map functions:

    • kept state resumes and the radio still maps;
    • a changed managed TTBR, a changed TCR on another stream, and a DART that reads back as reset each make resume return -EIO with the radio streams disabled, managed_failed and every borrowed consumer failed set, and a later 16 KiB radio map refused;
    • a negative control without the quarantine fails.

    This checks the failure path on the host, not physical retention on J713; the retained path is unchanged.

• Checks: apple-dart.o builds with W=1 and no warnings. checkpatch --strict is clean (apart from the MAINTAINERS note for the new test file), dt_binding_check passes for both schemas, and the J713 dtbs_check adds nothing new.

#432 (cpuidle) is what lets the cores reach deep idle while asleep.

[RyanTheTide]

The J713 suspend and lid results in the PR cover the normal retained-state path, including display, GPU, radio and audio after resume. The new retained-DART admission flags are already restricted to J713 by probe, and the PMGR/mailbox opt-ins are set only in its DT.

The lost-retention branch still needs a fix. apple_dart_resume() returns -EIO when the saved TCR/TTBR state changed, but leaves the managed-radio and borrowed-display admission state usable. A resume callback error does not stop the PM core from resuming consumers. Please quarantine the affected streams/session before returning, using the existing managed/borrowed failure paths as appropriate, and add a host fault case that changes a retained register and verifies later consumer operations are refused. Logging the loss alone does not establish a safe resume boundary.

The host fault case against the extracted production functions changed a live TTBR. Resume returned -EIO, managed_failed stayed false, and a later 16 KiB map succeeded. The control that sets the failure flag refused that map. This checks the managed path on the host, not physical J713 retention.

This head also needs a rebase onto 3b77f0f97eb0. Commit 5afbdfeb0c79 conflicts in dcp_platform_suspend() with the base's aborted-takeover guard. The local test assembly keeps that guard before the new M4 suspend call.

The stacked kernel at 5b6f5515e2f8 passed both full universal and Fedora-base WERROR builds (Image, modules and DTBs); both preserve all 131 common-config settings. ARM64 16 KiB KUnit passed 275/275. DCP host tests passed 63/63, and the PMGR and radio SID checks passed. The M4 host suite has the same five failing groups as the pinned base (18/23 on each); the AFKv2 HPD harness also fails on both. These are recorded failures, not a full host-suite pass. All six T8132 DTBs were checked: 13 schema diagnostics against 11 on the base, with exactly two new DCP dma-noncoherent binding errors from #429. The seven targeted binding checks passed. These software gates do not replace the required shared-code boots.

This also changes shared DART save-state code, PMGR and mailbox PM, so the assembled kernel needs Neo and M1 Pro or M2 regression boots with s2idle, display/GPU, radio and audio checks. Merge after #429/#432 for the qualified J713 graphics and idle combination. #450 follows this and #434.

[erikwb]

Updated (4fe8d2f), rebased onto 3b77f0f97eb0.

Lost retention: apple_dart_resume() now goes through apple_dart_resume_retained() for a retained DART. If the saved TCR/TTBR state changed, it quarantines through the existing translation-fault paths before returning -EIO:

• apple_dart_managed_quarantine() for the managed radio streams: managed_failed is set and their DMA disabled;
• apple_dart_quarantine_borrowed() for the DCP and its frontends.

Host fault case: tools/testing/selftests/drivers/apple/dart-s2idle-retention.py runs on the extracted production apple_dart_state_retained(), the two quarantine helpers, apple_dart_resume_retained() and apple_dart_map_pages().

• Each of these makes resume return -EIO, with managed_failed and every borrowed failed set, the radio SIDs disabled, and a later 16 KiB map refused:
  • a changed managed TTBR;
  • a changed TCR on another stream of the same DART;
  • a DART that reads back as reset.
• The retained case still resumes and maps.
• A negative control with the quarantine removed fails at runtime.
• On the display side, m4-dcp-errors already checks that a failed DCP domain refuses apple_dart_dcp_map_owned() with -EIO.

As you said, this checks the failure path on the host, not physical retention.

Rebase: commit 3 now keeps the aborted-takeover guard before the M4 suspend call. Range-diff against the old head: commits 1-2 and 5-10 are identical, commit 3 differs only by that context, and commit 4 has the quarantine and test.

Checks:

• apple-dart.o builds with W=1 and no warnings; checkpatch --strict is clean.
• The apple-dcp host suite is 57/63, with the same six failures as the bare base; m4-dcp-errors and the radio SID checks pass.

Noted on the Neo and M1 Pro/M2 regression boots; I don't have those machines.

[erikwb]

Superseded by #830 and #831, both on integration/aurora, with the commits unchanged:

• drm/apple: m4: Power the display down for system sleep, while blanked and while static #830: the DCP's system-sleep handshake (commits 1–3), now together with J713: power the display down while it is blanked (depends on #449) #501 and J713: cache a static frame while nothing is flipping (depends on #501) #503's display power commits.
• J713 (M4 MacBook Air): suspend-to-idle #831: the iommu, PCI, pmdomain and mailbox pieces and the J713 DT. The lost-retention quarantine and its host test are there, and so is your request for Neo and M1 Pro/M2 regression boots.